AXOlotlvaNschmozelot/keep-open-weights-alive

Keep your local model library alive

Part 4 of The Open-Weights Lifecycle — acquire → verify → run → maintain → retire.

You downloaded the models (Part 1), sized them (Part 2), and ran them (Part 3). But an archive is not a "set it and forget it" thing. Files rot, disks die, and future-you forgets what half the folders even are. Keeping a library alive is a little bit of light maintenance, done on a rhythm — not a big project.

Here are the five habits that keep a local model archive trustworthy over years.

1. Re-verify — catch silent corruption before it catches you

Bits rot. A drive can flip a byte with no error message, and a corrupted model file may load and simply behave slightly wrong — the worst kind of failure, because you won't notice. The cure is cheap: when you first archive a model (Part 1), save its checksums; then re-hash the files every so often and compare.

# stored at acquire time:  model.safetensors  ->  <hash-A>
# months later, re-check:
sha256sum model.safetensors        # Linux/macOS
Get-FileHash model.safetensors     # Windows PowerShell
# same hash  = intact.   different hash = corruption, restore from your second copy.

Put it on a calendar: a full re-verify once or twice a year is plenty for cold storage.

2. Redundancy — one copy is not a backup

The old rule still holds: 3-2-1 — three copies, on two kinds of media, one of them off-site (or at least physically separate). A single external drive is one power surge away from taking your whole library with it. At minimum, keep a second copy of the masters (the originals you can't re-derive); the quants you can always rebuild.

3. Know what's replaceable vs irreplaceable

Not everything in the folder is equally precious:

• The master (the full-precision weights) is the irreplaceable original — protect it hardest.
• Quants (GGUF Q4, Q8, …) are derived — you can regenerate any of them from the master. If space is tight, thin out the quants, not the master.

Deduplicate on that basis: keeping the master plus one working quant is usually enough. Five overlapping quants of a model you rarely run is just disk you're maintaining for nothing.

4. Keep a plain-text index

The most valuable file in the whole archive costs nothing: a simple list of what you have, where it lives, its checksum, and when you last verified it. Future-you (or a fresh machine after a rebuild) can read it at a glance. No database, no tooling — a text file that outlives every app.

5. Refresh the media

Cold drives are not forever. Unpowered hard disks and SSDs both degrade over years — SSDs can lose data if left without power for a very long time; disks develop bad sectors. Power your archive drive on periodically (which pairs nicely with the re-verify in habit 1), and plan to migrate to fresh media every few years. The data is only as durable as the thing it sits on.

When a re-verify fails — don't panic, it's fixable

So you re-hashed a file and the number is different. First, the important part: a changed hash means that copy is damaged — not that the model is lost. Here's the ladder, calmest step first.

1. Rule out a fluke. A single bad read, a flaky USB cable, or a drive that hadn't fully spun up can produce a wrong hash. Re-run the check; try another cable or port; unplug and replug. Often the second read matches and there was never a real problem.

2. Pin down which file. Re-hash just the suspect so you know exactly what's damaged:

Get-FileHash model-00003-of-00007.safetensors      # Windows
sha256sum   model-00003-of-00007.safetensors       # Linux/macOS

3. Restore it from your second copy. This is the whole reason habit 2 (3-2-1) exists. Copy the good file over the bad one, then re-hash — it matches the manifest again, and you're done:

robocopy E:\copy2\<model> D:\<model> model-00003-of-00007.safetensors   # Windows, one file
cp /copy2/<model>/model-00003-*.safetensors  ~/<model>/                 # Linux/macOS

4. No second copy? Re-download just what broke. If the model is still on the hub, fetch a fresh copy and re-verify — the downloader only re-pulls what's missing or changed:

hf download <repo-id> --include "model-00003-of-00007.safetensors" --local-dir ./<model>

5. Damaged, no second copy, and no longer online? Then it's genuinely lost — which is exactly the case habits 1 and 2 exist to prevent. You re-verify often precisely so you catch rot while a good copy still exists somewhere. (This is also the deepest reason to keep, not casually delete, rare models — see Part 5.)

Finally, ask why it rotted. One bad file can be bad luck; several point to a failing drive. Check its health, and if it's dying, migrate everything off it now:

Get-PhysicalDisk | Select FriendlyName, HealthStatus, OperationalStatus   # Windows
smartctl -H /dev/sdX                                                       # Linux (smartmontools)

The reassuring version: a mismatch is a drill you already prepared for. With a second copy or a live hub link, it's a two-minute fix. Without either, it's the warning that makes you glad you kept reading.

The point

None of this is heavy. Twice a year: plug the drive in, re-hash, glance at the index, note anything that changed. That's the difference between a pile of files that might still be good and a library you can actually trust the day you need it — which, after all, was the whole reason to keep your own copy.

Cheat-sheet

# Save a manifest at acquire time (do this ONCE, keep the file with the model):
Get-FileHash *.safetensors > manifest.txt        # Windows PowerShell
sha256sum   *.safetensors > manifest.txt          # Linux/macOS

# Re-verify months later — compare against that saved manifest:
Get-FileHash *.safetensors                         # Windows (eyeball vs manifest)
sha256sum -c manifest.txt                           # Linux/macOS — auto-compares, prints OK / FAILED

# A file failed? Restore it from your second copy:
robocopy <copy2-dir> <live-dir> <filename>          # Windows, single file

# ...or re-fetch just the broken file from the hub:
hf download <repo-id> --include "<filename>" --local-dir <dir>

# Check the drive's health (is it a fluke, or a dying disk?):
Get-PhysicalDisk | Select FriendlyName, HealthStatus     # Windows
smartctl -H /dev/sdX                                      # Linux (smartmontools)

sha256sum -c manifest.txt is the friendliest of these — it reads your saved manifest and tells you, file by file, OK or FAILED. On Windows there's no built-in -c, so you compare Get-FileHash output against the manifest yourself (or script it).

---

Part 3 — Run it: run your first local model in 5 minutes. Next, the finale — Let it go: retiring a model you no longer need, without losing the option to get it back.